CONCEPT
Frozen Accidents in AI
Jacob’s evolutionary principle applied to machine learning: the design choices that seemed reasonable when AI systems were small have become permanently load-bearing—too entangled with everything built on top of them to change without collapsing the edifice.
The recurrent laryngeal nerve in the giraffe travels several meters down the neck and back up to reach a larynx inches from where it started. No engineer would route a nerve this way. A tinkerer would, because in the fish ancestor the path was direct and evolution could only stretch it, never reroute it. This is François Jacob's concept of the frozen accident: a choice that made sense at the moment it was made, that everything subsequent was built on top of, and that now cannot be changed without rebuilding everything above it—which is never worth the cost. Jacob used it to explain why living organisms carry absurdities alongside their brilliance; the absurdities are the fossils of a process that could only modify, never replace. AI systems are developing a fossil record of their own. The byte-pair tokenizer, chosen early because it was a reasonable compromise for the hardware and data of the moment, has become the frozen
Keep reading with YOU ON AI
Unlock the full book, 10,000+ field-guide entries, and a 1000+ thinker library. If you have a book code, register now — it takes a minute.